Commit
·
b52bed7
1
Parent(s):
2d191f6
Added inference
Browse files- app.py +54 -1
- requirements.txt +2 -0
app.py
CHANGED
@@ -7,6 +7,7 @@ import torch.nn as nn
|
|
7 |
from model import Projections
|
8 |
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
9 |
import gradio as gr
|
|
|
10 |
|
11 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
12 |
projections = Projections(512, 3072)
|
@@ -47,7 +48,59 @@ whisper_processor = WhisperProcessor.from_pretrained(whisper_model_name)
|
|
47 |
whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_name)
|
48 |
|
49 |
def infer(message, history):
|
50 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
|
52 |
examples=[{'text':"I am planning to buy a dog and a cat. Suggest some breeds that get along with each other"},
|
53 |
{'text':"Explain biased coin flip"},
|
|
|
7 |
from model import Projections
|
8 |
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
9 |
import gradio as gr
|
10 |
+
import librosa
|
11 |
|
12 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
13 |
projections = Projections(512, 3072)
|
|
|
48 |
whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_name)
|
49 |
|
50 |
def infer(message, history):
|
51 |
+
max_generate_length = 100
|
52 |
+
combined_embeds = []
|
53 |
+
|
54 |
+
with torch.no_grad():
|
55 |
+
if message['file']:
|
56 |
+
projected_image_embeds = None
|
57 |
+
audio_text_embeds = None
|
58 |
+
for path in message['file']:
|
59 |
+
|
60 |
+
if path.endswith(('.jpg', '.png', '.jpeg')):
|
61 |
+
image = clip_preprocess(Image.open(path)).unsqueeze(0).to(device)
|
62 |
+
image_features = clip_model.encode_image(image)
|
63 |
+
projected_image_embeds = projections(image_features.to(torch.bfloat16)).unsqueeze(0)
|
64 |
+
|
65 |
+
elif path.endswith(('.mp3', '.wav')):
|
66 |
+
# Load and preprocess the audio
|
67 |
+
speech, rate = librosa.load(path, sr=16000)
|
68 |
+
input_features = whisper_processor(speech, return_tensors="pt", sampling_rate=16000).input_features
|
69 |
+
predicted_ids = whisper_model.generate(input_features)
|
70 |
+
transcription = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)
|
71 |
+
prompt = tokenizer.apply_chat_template([{"from": "human", "value": transcription}], tokenize=False, add_generation_prompt=True)
|
72 |
+
prompt_tokens = tokenizer(prompt, padding=True, truncation=True, max_length=2048, return_tensors="pt")['input_ids']
|
73 |
+
audio_text_embeds = model.get_input_embeddings()(prompt_tokens)
|
74 |
+
|
75 |
+
if projected_image_embeds:
|
76 |
+
combined_embeds.append(projected_image_embeds)
|
77 |
+
|
78 |
+
if audio_text_embeds:
|
79 |
+
combined_embeds.append(audio_text_embeds)
|
80 |
+
|
81 |
+
if message['text']:
|
82 |
+
prompt = tokenizer.apply_chat_template([{"from": "human", "value": transcription}], tokenize=False, add_generation_prompt=True)
|
83 |
+
prompt_tokens = tokenizer(prompt, padding=True, truncation=True, max_length=2048, return_tensors="pt")['input_ids']
|
84 |
+
text_embeds = model.get_input_embeddings()(prompt_tokens)
|
85 |
+
combined_embeds.append(text_embeds)
|
86 |
+
|
87 |
+
combined_embeds = torch.cat(combined_embeds,dim=1)
|
88 |
+
|
89 |
+
#val_combined_embeds = torch.cat([val_image_embeds, img_token_embeds, val_q_embeds], dim=1) # 4, 69, 2560
|
90 |
+
predicted_caption = torch.full((1,max_generate_length),50256).to(device)
|
91 |
+
|
92 |
+
for g in range(max_generate_length):
|
93 |
+
phi_output_logits = model(inputs_embeds=combined_embeds)['logits'] # 4, 69, 51200
|
94 |
+
predicted_word_token_logits = phi_output_logits[:, -1, :].unsqueeze(1) # 4,1,51200
|
95 |
+
predicted_word_token = torch.argmax(predicted_word_token_logits, dim = -1) # 4,1
|
96 |
+
predicted_caption[:,g] = predicted_word_token.view(1,-1)
|
97 |
+
next_token_embeds = model.get_input_embeddings()(prompt_tokens) # 4,1,2560
|
98 |
+
combined_embeds = torch.cat([combined_embeds, next_token_embeds], dim=1)
|
99 |
+
|
100 |
+
predicted_captions_decoded = tokenizer.batch_decode(predicted_caption,ignore_index = 50256)[0]
|
101 |
+
|
102 |
+
return predicted_captions_decoded
|
103 |
+
|
104 |
|
105 |
examples=[{'text':"I am planning to buy a dog and a cat. Suggest some breeds that get along with each other"},
|
106 |
{'text':"Explain biased coin flip"},
|
requirements.txt
CHANGED
@@ -3,6 +3,8 @@ clip @ git+https://github.com/openai/CLIP.git@dcba3cb2e2827b402d2701e7e1c7d9fed8
|
|
3 |
colorama==0.4.6
|
4 |
datasets==3.0.0
|
5 |
dill==0.3.8
|
|
|
|
|
6 |
multiprocess==0.70.16
|
7 |
numpy==1.26.4
|
8 |
pandas==2.2.2
|
|
|
3 |
colorama==0.4.6
|
4 |
datasets==3.0.0
|
5 |
dill==0.3.8
|
6 |
+
gradio==5.0.2
|
7 |
+
librosa==0.10.2
|
8 |
multiprocess==0.70.16
|
9 |
numpy==1.26.4
|
10 |
pandas==2.2.2
|